Memory Court is an auditable autonomous-agent demo built for OpenAI Build Week. GPT-5.6 investigates a memory case, chooses one structured action per turn, and proposes bounded cognitive-state interventions. sonuv-guard independently adjudicates every intervention as COMMIT, REPAIR, REJECT, or FORGET before any state is applied.

The interface makes the boundary explicit: GPT-5.6 supplies the investigation and proposal; Guard controls structured state execution. Guard does not review or certify the model's natural-language reasoning.

• Application: https://memory-court-build-week.vercel.app
• API health: https://memory-court-api-production.up.railway.app/api/health
• Source: https://github.com/billgaohub/memory-court-build-week

The public build keeps the server-side GPT-5.6 Responses API path intact. Its current key has no billable quota, so the default judgeable artifact is a competition-period trace generated by GPT-5.6 Sol inside Codex, executed against the real Guard adapter, and displayed as REPLAY MODE. It is explicitly not presented as an API-live run.

1. Open the application and select The Silent Lifeboat, the competition-period case.
2. Inspect the gold REPLAY MODE trace. Its provenance banner states that GPT-5.6 Sol generated the actions inside Codex—not through the OpenAI API—and the checked-in source action file is reproducibly adjudicated by sonuv-guard.
3. If the server later has usable API quota, choose Run live audit. Otherwise the failed live request automatically returns to the same labeled replay.
4. Choose Export audit JSON to inspect the complete evidence artifact.

No account is required. The public demo is free to test and rate-limits each client to five new live sessions per ten minutes.

Vercel: React + TypeScript + Vite
             |
             v
Railway: FastAPI -> OpenAI Responses API (GPT-5.6)
             |
             v
       schema validation -> sonuv-guard -> committed state + audit event

The live loop is deliberately bounded to 8 model calls, 8 events, 3 intervention proposals, and 600 output tokens per call. Each OpenAI request has a 30-second timeout; timeout and rate-limit responses retry once. Sessions live in one API process for one hour. This is a transparent hackathon-grade design, not a distributed production control plane.

Requirements: Python 3.11+, Node.js 22+, and an optional OpenAI API key for live mode.

git clone https://github.com/billgaohub/memory-court-build-week.git
cd memory-court-build-week

python3 -m venv .venv
. .venv/bin/activate
pip install -e 'backend[test]'

export OPENAI_API_KEY='your server-side key'
export ALLOWED_ORIGINS='http://localhost:5173'
uvicorn memory_court.app:app --app-dir backend --host 127.0.0.1 --port 8000

In a second terminal:

cd frontend
npm ci
VITE_API_BASE_URL=http://localhost:8000 npm run dev

Without OPENAI_API_KEY, /api/health reports live_available=false and the frontend presents the labeled replay.

bash scripts/verify.sh

The gate runs 36 backend tests, including exact reproduction of the Codex-generated Guard trace, frontend interaction tests, TypeScript checking, a production build, vendored-source hashes, and—inside the original source workspace—the upstream 121-test sonuv-guard regression suite.

The competition-period Silent Lifeboat replay was generated by GPT-5.6 Sol inside Codex in primary task 019f725e-6f43-78c2-8587-4ad6a3725d9f. The exact six source envelopes are checked in at replay/silent_lifeboat.codex-actions.json. A backend test executes them through the same AgentSession and sonuv-guard adapter and requires the published six-event replay, including REPAIR followed by COMMIT, to match exactly.

Separately, the retained live backend uses the async OpenAI Python SDK Responses API with gpt-5.6, Pydantic structured outputs, low reasoning effort, low verbosity, store=false, and a 600-token ceiling. Each response must be exactly one of:

• inspect_memory
• propose_intervention
• finalize

Only propose_intervention reaches Guard. Invalid fields, booleans masquerading as integers, out-of-range values, and empty patches are rejected before adjudication.

The recorded Codex trace and API-live path are never conflated: replay uses mode=replay, api_live=false, and the model label gpt-5.6-sol via Codex (recorded).

Codex was used to review the original game plan, isolate a standalone architecture, create the approved design specification, implement the backend and frontend test-first, package the pre-existing Guard runtime with provenance hashes, prepare deployment, and run the completion audit. The primary build task ID and commit evidence are recorded in CODEX_EVIDENCE.md.

The autonomous loop, FastAPI service, React audit console, Silent Lifeboat case, replay system, deployment packaging, tests, and submission materials were created during the competition period. A precise snapshot of pre-existing sonuv-guard commit 62157c5 and the pre-existing Last Birthday case are disclosed in PREEXISTING_VS_NEW.md.

• Railway builds backend/Dockerfile from the repository root. Configure OPENAI_API_KEY, OPENAI_MODEL=gpt-5.6, TRUST_PROXY=true, and ALLOWED_ORIGINS.
• Vercel uses frontend as Root Directory, npm run build as Build Command, dist as Output Directory, and the Railway origin as VITE_API_BASE_URL.
• After Vercel assigns the production domain, set that exact origin in Railway's ALLOWED_ORIGINS and redeploy.

See SECURITY.md for the threat boundary and SUBMISSION.md for competition copy.

MIT. See LICENSE.